Looking at the Overlooked: An Analysis on the Word-Overlap Bias in Natural Language Inference